Financial Prediction Study Based On Discrete Bayesian Network

With the global economy developing constantly,to stand out in the fierce and intense competition,the attitude towards risk of companies tends to change into the type of preference.Thus,if companies fail to choose the appropriate developing strategies,it will be possible for them to be trapped in financial crisis,and this will definitely result in the loss of benefit stakeholders.Capital market in Hong Kong is gradually becoming the first choice of entering the international markets.But,due to the influence of risk factors in the global financial market,plus the high liquidity of capital,for the companies themselves,the third-party financial institutions and investors,it's important to have a general understanding of the financial position of the listed companies in Hong Kong by appropriate financial prediction method.Based on the analysis of the current study status of financial prediction method at home and abroad,this paper chooses the method of Bayesian network which is more suitable for the application background of financial prediction.Bayesian network has been applied to various of areas,and during the process of application,we need to solve two important problems,that is,the method of index selection and the method of handling continuous variables.But in the past study,the further studies about the two problems are relatively few.This paper selects the study data of the listed companies in 2013 from the database of Guo Tai An,and predicts the financial condition of corresponding companies in 2014.For the first problem,this paper selects the heuristic method of index selection,builds six different models having different correlation coefficients thresholds and discusses the difference of prediction result of these models.For the second problem,this paper handles the discrete variables by the discretion methods based on distribution characteristics and compares the different discretion methods.During the process,this paper tries to seek for the better discretion method for variables corresponding to different correlation coefficients.By constant comparison and analysis,this paper comes to the following conclusions.Firstly,when the correlation coefficient threshold is 0,the prediction effect of model is the best no matter which discretion method is adopted,and this indicates that the indexes having low correlation coefficients with the purpose node can help increase the prediction accuracy.Secondly,from the angle of single model,generally,the E-SM method is better than the E-PT method,and the two methods are significantly prior to the other four discretion methods.Thirdly,from the angle of contacted models,for the indexes having different correlation coefficients with the purpose node,we can select the best discretion method.Finally,after the appropriate discretion of every index,the entire prediction effect has been significantly promoted,and this indicates that the discretion method is reasonable and reliable.